openstack team mailing list archive

["Day, Phil" <philip.day@xxxxxx>]

> However I'd like to point out that the math below is misleading (the average time for the non-blocking case is also miscalculated but 
> it's not my point). The number that matters more in real life is throughput. For the blocking case it's 3/30 = 0.1 request per second.

I think it depends on whether you are trying to characterise system performance (processing time) or perceived user experience (queuing time + processing time).   My users are kind of selfish in that they don't care how many transactions per second I can get through,  just how long it takes for them to get a response from when they submit the request.

Making the DB calls non-blocking does help a very small bit in driving up API server utilisation  - but my point was that time spent in the DB is such a small part of the total time in the API server that it's not the thing that needs to be optimised first.     

Any queuing system will explode when its utilisation approaches 100%, blocking or not.   Moving to non-blocking just means that you can hit 100% utilisation in the API server with 2 concurrent requests instead of *only* being able to hit 90+% with one transition.   That's not a great leap forward in my perception.

Phil

-----Original Message-----
From: Yun Mao [mailto:yunmao@xxxxxxxxx] 
Sent: 03 March 2012 01:11
To: Day, Phil
Cc: openstack@xxxxxxxxxxxxxxxxxxx
Subject: Re: [Openstack] eventlet weirdness

First I agree that having blocking DB calls is no big deal given the way Nova uses mysql and reasonably powerful db server hardware.

However I'd like to point out that the math below is misleading (the average time for the nonblocking case is also miscalculated but it's not my point). The number that matters more in real life is throughput. For the blocking case it's 3/30 = 0.1 request per second.
For the non-blocking case it's 3/27=0.11 requests per second. That means if there is a request coming in every 9 seconds constantly, the blocking system will eventually explode but the nonblocking system can still handle it. Therefore, the non-blocking one should be preferred.
Thanks,

Yun

>
> For example in the API server (before we made it properly multi-threaded) with blocking db calls the server was essentially a serial processing queue - each request was fully processed before the next.  With non-blocking db calls we got a lot more apparent concurrencybut only at the expense of making all of the requests equally bad.
>
> Consider a request takes 10 seconds, where after 5 seconds there is a call to the DB which takes 1 second, and three are started at the same time:
>
> Blocking:
> 0 - Request 1 starts
> 10 - Request 1 completes, request 2 starts
> 20 - Request 2 completes, request 3 starts
> 30 - Request 3 competes
> Request 1 completes in 10 seconds
> Request 2 completes in 20 seconds
> Request 3 completes in 30 seconds
> Ave time: 20 sec
>
>
> Non-blocking
> 0 - Request 1 Starts
> 5 - Request 1 gets to db call, request 2 starts
> 10 - Request 2 gets to db call, request 3 starts
> 15 - Request 3 gets to db call, request 1 resumes
> 19 - Request 1 completes, request 2 resumes
> 23 - Request 2 completes,  request 3 resumes
> 27 - Request 3 completes
>
> Request 1 completes in 19 seconds  (+ 9 seconds) Request 2 completes 
> in 24 seconds (+ 4 seconds) Request 3 completes in 27 seconds (- 3 
> seconds) Ave time: 20 sec
>
> So instead of worrying about making db calls non-blocking we've been working to make certain eventlets non-blocking - i.e. add sleep(0) calls to long running iteration loops - which IMO has a much bigger impact on the performance of the apparent latency of the system.>>>> Thanks for the explanation. Let me see if I understand this.